Schedule-Based Robotic Search for Multiple Residents in a Retirement Home Environment

Markus Sebastian Schwenk, Tiago Stegun Vaquero, Goldie Nejat, Kai Oliver ArrasAAAI 2014 · DOI

Introduction

Social isolation in retirement homes is a critical health concern, yet staff shortages often make it difficult to organize group activities. While socially assistive robots can help by rounding up residents for events, finding people who are constantly moving is a complex "needle in a haystack" problem. Traditional search algorithms often fail to account for the structured nature of human behavior in institutional settings. To address this, researchers have developed M-USB (Multi-User Schedule Based) Search, a planning approach that leverages residents' daily schedules and past observations to optimize multi-target search paths.

The M-USB Search Framework

The M-USB framework models the search task as a finite-horizon Markov Decision Process (MDP). The robot's goal is to maximize the expected number of residents found within a specific timeframe. Central to this approach is the calculation of "rewards" for visiting specific locations based on spatio-temporal likelihood functions.

The system integrates four primary data sources to maintain a probability map for each resident:

  1. Schedule Analyzer: Utilizes official daily appointments.
  2. Last Known Location: Accounts for the time elapsed since the last sighting and potential travel distance.
  3. Learned Behavior: Incorporates long-term individual patterns (e.g., typical walking times).
  4. Environment Topology: Weights common areas and private rooms more heavily than transit zones like hallways.
Figure 1: Simulation environment setup and the dynamic weighting/priority functions used by the robot.

A key innovation is Dynamic Weighting. The robot adjusts the influence of each source based on its variance (certainty). For instance, during a scheduled event, the "Schedule" weight increases due to higher certainty, whereas "Learned Behavior" or "Environment" weights may dominate during unstructured time. The robot uses backwards induction to find the optimal sequence of actions—drive, rest, or search—to maximize its success rate before a deadline.

Simulation and Evaluation

The researchers evaluated M-USB in a custom simulation environment modeled after a real retirement home floor plan, featuring 25 private rooms and various common areas. The simulation included 25 residents with unique, realistic schedules. To test robustness, a "miss probability" ($p_{miss} = 0.1$) was introduced, simulating residents who occasionally deviate from their routines to wander elsewhere.

Key Results

M-USB Search was compared against two baselines: a Random Walk and a Weighted Informed Walk (which understands room types but lacks schedule data).

Figure 6: Comparison of the number of regions visited to find a target number of residents.

The results demonstrated that M-USB is significantly more efficient. As shown in the performance metrics, the robot required far fewer room visits to locate the same number of residents compared to the baseline methods. This efficiency translates to reduced "robot traffic" in the facility and improved battery life. Furthermore, the system showed linear scalability, indicating its suitability for larger facilities with more residents.

Conclusion and Future Work

M-USB Search proves that treating digital schedules as "virtual sensors" can drastically improve robotic search efficiency in structured environments. While currently validated in simulation, future work will focus on real-world deployment, addressing challenges like hallway clutter and real-time schedule updates. Additionally, incorporating social nuances—such as the cost of interrupting specific activities—could further refine the robot's decision-making process.

Summary updated

BibTeX
@inproceedings{schwenk2014schedule,
  author = {Markus Sebastian Schwenk and Tiago Stegun Vaquero and Goldie Nejat and Kai Oliver Arras},
  title = {{Schedule-Based Robotic Search for Multiple Residents in a Retirement Home Environment}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2014},
  doi = {10.1609/aaai.v28i1.9050},
}
Figure